Refrax
github.com
- Free plan
- Windows
- macOS
- Linux
What Refrax is
Refrax is a free desktop tool for refactoring Java code with an AI-powered agent implemented in Go. It divides a broad code-polishing request into smaller tasks for agents such as a critic, reviewer, and editor, then runs those tasks autonomously. Multiple language models communicate through the A2A protocol. The README lists DeepSeek and OpenAI as providers, while the latest listed release also supports custom providers by URL. Pre-built binaries are available for Linux, macOS, and Windows; building from source requires Go 1.24.1 or later. Some operations need an AI provider API token, supplied through a command-line flag, environment variable, or .env file. Refrax does not inherently verify that applied changes are correct or harmless, but reviewer commands such as Maven tests can report feedback. The CLI prints interaction statistics or exports them as CSV. It is MIT-licensed and described as an early prototype subject to rapid change.
Who it suits
Refrax suits Java developers who want autonomous, batch-oriented refactoring across Linux, macOS, or Windows. It is also relevant to teams able to configure provider tokens and reviewer checks.
Good to know, and what to watch
- Free plan and MIT license
- Pre-built binaries for three desktop platforms
- Breaks refactoring into specialist agent tasks
- Supports DeepSeek and OpenAI providers
- Exports interaction statistics as CSV
- Early prototype subject to rapid change
- Does not inherently verify applied changes
- Some operations require an AI provider token
- Source installation requires Go 1.24.1 or later
The verdict
Refrax combines autonomous Java refactoring with multiple model providers and configurable review commands. Its prototype status and lack of inherent correctness checks make validation an important part of use.
What the maker publishes
- Agent protocol
- Refrax integrates multiple language models that communicate using the A2A protocol. source
- AI providers
- The README lists DeepSeek and OpenAI as supported providers; the latest listed release also adds custom providers by URL. source
- Authentication
- Some operations require an AI provider API token, supplied by command-line flag, environment variable, or a .env file. source
- Authentication
- Some operations require an AI provider API token, which can be supplied with a command-line flag, environment variable, or .env file. source
- Autonomy
- The README says Refrax can run autonomously for minutes or hours after it is started. source
- Installation
- The project offers pre-built binaries for macOS, Windows, and Linux, and installation from source requires Go 1.24.1 or later. source
- Installation
- The project offers prebuilt binaries for macOS, Windows, and Linux, and source installation requires Go 1.24.1 or later. source
- License
- The repository states that Refrax is licensed under the MIT License. source
- License and maturity
- The repository is MIT-licensed and labels Refrax an early prototype subject to rapid change. source
- Maturity
- The README describes Refrax as an early prototype subject to rapid change. source
- Protocol
- Refrax connects multiple large language models using the A2A protocol. source
- Providers
- The README lists DeepSeek and OpenAI as supported AI providers. source
- Purpose
- Refrax is an AI-powered agent for refactoring Java code, implemented in Go. source
- Purpose
- Refrax is an AI-powered agent for refactoring Java code, implemented in Go. source
- Results limitation
- Refrax does not inherently verify that its applied changes are correct or cause no issues. source
- Reviewer checks
- Users can pass one or more commands with --check for the reviewer agent to run and report feedback to the facilitator. source
- Statistics
- The refactor command can export interaction statistics in CSV format or print them to the console. source
- Usage statistics
- The CLI can output interaction statistics to the console or to a CSV file. source
- Validation
- Refrax does not inherently verify whether changes are correct; users can configure reviewer checks with commands such as Maven tests. source
- Workflow
- It breaks large code-polishing requests into smaller tasks for agents such as a critic, reviewer, and editor, and runs them autonomously. source
- Workflow
- It breaks larger code-polishing requests into smaller tasks for agents such as a critic, reviewer, editor, and facilitator. source
Checked against the maker's pages on 3 Oct 2026.